Posts

• Pythia-14M on four rented GPUs and a bucket

  A Salad training run, including failures and comparisons with the published checkpoints.

• MLX and PyTorch learners in one run

  What crosses the mailbox, what has to match between the two models, and what the library checks.

• Merging deltas: RDA against a weighted mean

  How contributions are weighted, the two merges on small inputs, and how the choice changes the outer step.

• Implementing a mailbox backend

  The four methods, the guarantees a store must give, and a filesystem backend tested against the suite.